GHOST: Geometry-Hierarchical Online Streaming Token Eviction for Efficient 3D Reconstruction

Researchers introduced GHOST, a training-free KV-cache management framework for streaming 3D reconstruction from monocular video.

GHOST uses geometry-aware importance scoring, token protection, and layer-wise budget allocation to evict redundant tokens during long sequences.

The method nearly halves cache usage and speeds up inference by 1.75x without major quality loss.

It addresses a key memory bottleneck, making real-time and large-scale 3D reconstruction more practical.